Artificial Intelligence, Personalized Persuasion, and Climate Attitudes

Elena Pro, António Valentim

European Institute, London School of Economics

Climate Information and Climate Concern

The Problem: Persuading people to care about and act toward climate change is still one of the major collective challenges we face as a society

Effective communication is crucial
- Over $15 million spent in science-based public information campaigns (United Nations 2025)

However, effects are small and short-lived (Deryugina and Shurchkov 2016; van der Linden 2017)

In Practice:

Climate change image

80%

care about climate change

BUT

don’t always know how to address it

Why is this happening?

The issue:
- Psychological distance and intertemporal discounting reduce climate urgency (Trope & Liberman, 2010).
- Emotional disconnection from abstract risks (Weber, 2006)

The Problem 1/2

People see why climate change matters generally, but not why it matters to them

The (partial) Solution:

Tailored messaging aligned with individual beliefs and experiences is more effective (Goldberg and Gustafson 2025)

But traditional targeting faces critical limitations:
- Operates at group level (demographics, ideology)
- Expensive
- ⟹ Hard to scale

The Problem 2/2

How can we deliver personalized messages at scale, cost-effectively, and in a way that engages people on a personal level?

Our Solution

We argue that AI – and in particular its personalized content - could influence climate attitudes and overcome some of the limitations of pure information provision.

Key Advantages:

From a practical perspective, AI can:
- Scale: Online implementation + individual-level personalization
- Cost-effective: Reduce costs significantly (see Appendix for cost analysis)
From a theoretical perspective, AI can:
- Evidence-based reasoning (Costello et al., 2025)
- Non-judgmental interaction (Goel et al., 2024)
- Responsive dialogue (Santoro et al., 2024)

Research Design

Online Survey Experiment

Sample:
- N = 1,500-2,000 UK representative sample
  - Power analysis: 930 minimum for small-to-medium effect (d = 0.26-0.37)
- Recruited via established survey platform (Prolific) - the pipeline can be integrated into Qualtrics
- Context: UK offers valuable setting where climate skepticism and polarization is growing but not deeply entrenched

Survey Flow

AI diagram

See detailed prompts in the appendix Appendix

The Treatments

Participants are randomized into these conditions:

Control: Standard information provision about climate change
Non-Personalized AI: LLM conversation about climate change not tailored to the respondent
Personalized AI (Unrelated Political Topic): Personalized AI-driven conversation about a political topic unrelated to climate change
Personalized AI (Climate): Personalized AI-driven conversation about climate change

Treatments Explanation

Conditions 3 and 4 serve as placebos for each other to isolate the effect of personalization specifically in the context of climate change
Difference between personalized and non-personalized AI interactions is whether the AI is prompted to adapt its responses to the participant’s prior inputs (demographics, concerns, open-text answers)

An Example

AI diagram

Outcome Variables

Climate concern: 5-point scale from “Not at all concerned” to “Extremely concerned”
Pro-environmental behavior: Willingness to donate portion of $100 prize to climate NGO
Conversation spillover: Likelihood of discussing AI conversation with others
Political spillover: Choice to write message to Member of Parliament about climate change
Subjective climate beliefs: Open-ended question about personal beliefs on climate change + LLM-generated Likert scale

Timing and Analysis

Pre-post measurement enables both within-subject change detection and between-group comparisons (Change scores Δ = Post-treatment - Pre-treatment)
Quantiative Text Analysis of the conversations to unpack mechanisms
Follow-up survey to assess long-term effects